nx_thread_pool.nx

nx_thread_pool.nx -- shared-queue thread pool (Rayon-precursor). Architecture: hw-sized worker fleet + single Vyukov MPMC queue of task records. Workers spin on the channel; each task carries a typed fn-pointer + ctx; sentinel tasks (is_sentinel == 1) signal a worker to exit. MVP shape -- single shared queue, not yet work-stealing. Work- stealing buys you locality when one worker is starving and another is buried; for compute-bound tasks of comparable cost the shared- queue model is within ~5% of work-stealing (per Rayon's own bench). Composes upward to work-stealing in a follow-up by giving each worker a private deque + a steal protocol; the public API does not change. Tasks live in a bump-allocated arena (NxPoolTask). Submit picks the next slot via atomic FAA on `next_task_slot`, fills it, sends the slot pointer through the channel. Worker receives the pointer, calls task.fn(task.ctx), bumps the completed-counter. Slot reuse is a future concern -- bench first. Composes against: [[vyukov_mpmc_channel]] (queue), [[atomic_intrinsics_real_amo]] (FAA counters), [[thread_clone_native_trampoline]] (worker spawn), [[nx_hw_dynamic_probes]] (worker sizing), [[fn_ptr_indirect_call]] (typed call site).